vLLM vs ninfer for Qwen 3.8-27B on RTX 5090: a Reddit reality check
MaxKingCS · reddit · 2026-08-31
A Reddit user currently runs unsloth/Qwen3.8-27B-NVFP4 with 157k context via vLLM on an RTX 5090, and asks whether switching to the much-hyped ninfer is worth it. They also spotted a Hugging Face repo (gittensor-model-hub/Qwen3.8-27B-NVFP4-RTX5090) claiming more context and better performance than the unsloth build — unverified and suspicious in their view — and want optimal configs for use with agent harnesses like Hermes agent.
Related event: Single RTX 5090 Runs Qwen 27B at 200+ t/s with ninfer(2 posts)→
More from Infra
- Same 1,000 Cloudflare Durable Objects: $4,050/month cut to $22 with one trick — threepointone · 2026-10-16
- NVIDIA ships FP4-quantized Qwen3.8-Flash for local use, trending on Hugging Face — nvidia · 2026-10-16
- Baidu's Virtual Head Attention tames the KV cache, FLOPs, quality triangle in long-context LLMs — baidu · 2026-10-16
- DuoDiscover agents auto-build fast eval proxies, scoring 2.68x TPU kernel speedup — berkeley_ai · 2026-10-16
- Google Raises 2026 Capital Spending to Up to $205 Billion — emmanuelvivier · 2026-10-16
- Google open-sources AX, a Kubernetes-style orchestrator for billions of AI agent tasks — thisdudelikesAI · 2026-10-16